{"id":"4f542396-c7df-4068-ad26-9273215fa185","arxiv_id":"2608.08259","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A topological ML pipeline with a particle-level density filter recovers the intermittency index of a 5% CMC signal embedded in EPOS background in (eta, phi) space.","lead":"This paper shows that a two-stage topological machine learning pipeline can separate a weak injected critical signal from background in angular (eta, phi) particle distributions, and can recover the expected power-law scaling after filtering out background tracks. It is a Monte Carlo proof of concept that could eventually be applied to LHC data in the search for the QCD critical point.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported recovery is vulnerable to oracle tuning: epsilon_cut is selected after inspecting which cut reproduces the pure-CMC reference, so the 'accurate recovery' is a benchmark fit, not an out-of-sample measurement.","rationale":"The paper is a carefully constructed Monte Carlo benchmark: the periodic Delaunay filtration with phantom points, the Euler-characteristic shortcut for beta_1, the multiplicity/flow subtraction via azimuthal randomization, and the use of two independent classifiers are all coherent and are described in enough detail to reproduce. I am not questioning the internal consistency of the TDA construction; the concern is statistical. The final number that makes the paper's central claim — restoring phi2 from the diluted 5% mixture — is obtained after the threshold is chosen with knowledge of what the pure-CMC answer should be. That is acceptable as a closed-book benchmark only if the cut-selection step is validated out-of-sample or by an independent rule; as written, the 'recovery' is partly a constructor's result. The baseline inconsistency (0.73 vs 0.75) amplifies this because it is unclear which reference value the tuning targeted. This does not invalidate the paper as a proof-of-principle, but it does mean the advertised 'robust data-driven tool' claim is not yet supported. The reader's CONDITIONAL verdict is the right one, and my check is the one most likely to settle whether the condition is met.","tokens_in":20383,"tokens_out":5665,"duration_ms":51754,"concrete_test":"Hold out a validation set disjoint from the 5%-mixture test events. Fix epsilon_cut by an a priori rule using only training-event SHAP/gain attributions (e.g., the midpoint of the secondary |Δβ1| peak, or the scale maximizing the Δ-Betti dip separation from EPOS), without computing NFMs on CMC-reference or mixture events. Apply that fixed cut to held-out 5% mixture and pure-CMC events, compute phi2 with bootstrap uncertainties, and compare. In parallel, run the full event+particle selection chain on pure-EPOS events to test whether the filter alone generates a spurious power law. If the held-out phi2 deviates from the pure-CMC value by more than ~10%, or filtered EPOS yields a slope near 0.7, the two-stage recovery claim is not robust.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim — phi2 ≈ 0.73 recovered for the filtered 5% mixture against a pure-CMC reference of phi2 ≈ 0.75 — rests on the particle-level density threshold epsilon_cut = 0.02 rad. In Sec. VI.3 the text states that 'among these, the epsilon_cut ≥ 0.02 sample lies closest to the pure CMC reference', and Figs. 12–13 show that this is how the threshold was identified. The paper's stated justification that the cut is 'physically motivated by the feature attribution frameworks' does not remove the circularity: the SHAP/gain spectra show a broad feature across 0.02–0.05 rad, and the exact operating point is not fixed by a pre-specified rule before comparison with the reference. Thus the headline claim of 'successfully restores the power-law scaling ... enabling accurate recovery' is, as reported, a fit on the benchmark rather than an out-of-sample demonstration. The ambiguity is compounded by the pure-CMC baseline itself being quoted as phi2 ≈ 0.73 in Fig. 4 and as phi2 ≈ 0.75 in Sec. VI.3/Conclusions, and by the absence of any quoted uncertainty on the recovered indices.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript presents the first Critical Monte Carlo (CMC) based intermittency analysis in the two-dimensional angular (η,φ) phase space, using EPOS as the background model for Pb–Pb collisions at √sNN = 5.02 TeV. The authors embed CMC signal tracks at fractions λ = 1%, 3%, and 5%, show that standard normalized factorial moments (NFMs) are diluted by the background, and then propose a two-stage topological machine learning pipeline: event-level classification with TopoPointNet and XGBoost on Δ-Betti curves, followed by a particle-level nearest-neighbor density filter with threshold ε_cut. They report that this pipeline restores power-law scaling of NFMs and recovers an intermittency index φ2 ≈ 0.73 for the filtered 5% mixture, in close agreement with their pure CMC reference of φ2 ≈ 0.75. The paper also reports classification AUCs ~0.99, SHAP/gain feature attributions, and a discussion of limitations for experimental deployment.","tokens_in":20661,"tokens_out":3519,"duration_ms":36148,"significance":"If the recovery claim were validated out of sample, the paper would be a useful methodological contribution: it extends CMC-based intermittency studies from (p_x,p_y) to the experimentally more accessible (η,φ) plane, and it combines persistent homology, two independent classifiers, and explicit multiplicity-bias corrections. The periodic Delaunay construction with phantom points, the Euler-characteristic shortcut for β1, the use of azimuthal randomization, and the candid limitation paragraph are all strengths. However, the central quantitative claim is currently a benchmark fit rather than an out-of-sample measurement: the particle-level cut ε_cut is selected after inspecting which value best reproduces the pure CMC reference. Because the surviving tracks after such a density cut are mostly the injected CMC particles, the NFM recovery is to a substantial degree built into the construction. The paper needs a pre-specified threshold selection rule or an independent validation set, plus a consistent statement of the pure CMC reference value and its uncertainty.","major_comments":[{"comment":"The central recovery claim rests on a post hoc choice of ε_cut. The text states that among the tested thresholds, ‘the ε_cut ≥ 0.02 sample lies closest to the pure CMC reference,’ and Figs. 12–13 show that this is how the operating point was identified. Because the threshold is selected by closeness to the benchmark it is supposed to recover, the reported φ2 ≈ 0.73 is a fit on the benchmark, not an out-of-sample measurement. Please specify the threshold before any comparison with the pure CMC reference, or validate the threshold on a held-out set of mixtures and then apply it to the 5% sample without reference to the outcome.","section":"§VI.3 and Figs. 12–13"},{"comment":"The pure CMC reference value is quoted inconsistently: Fig. 4 gives φ2 ≈ 0.73, while §VI.3 and the Conclusions give φ2 ≈ 0.75, with no uncertainty on either. If the two values come from different fit ranges or different CMC samples, that must be stated; otherwise this is an internal inconsistency that directly affects the claimed 1% residual discrepancy between the recovered index and the reference.","section":"§II, Fig. 4 vs. §VI.3 and Conclusions"},{"comment":"The particle-level filter uses the same nearest-neighbor distance d_NN that defines the CMC clusters: after retaining tracks with d_NN ≤ 0.02 rad, the surviving sample is dominated by the injected CMC tracks by construction. This makes the NFM recovery partly mechanical, independent of the ML stage. Please quantify the post-filter signal purity and test a null control in which the same number of tracks is selected by a random or density-mismatched rule, to show that the restored power law is not an artifact of the cut itself.","section":"§VI.3 and Sec. IV.1"},{"comment":"The paper correctly notes that the theoretical value φ2 = 2/3 is derived for (p_x,p_y) space, not (η,φ), yet the CMC Lévy walk with μ = 1/6 is used directly in (η,φ) as a critical benchmark. The recovered index is therefore calibrated against a model-specific reference whose relation to QCD criticality in angular space is not independently established. This is a correctness-risk concern rather than a circularity claim; a concrete test would be to compare against the SCR model mentioned in the Conclusions, or to derive the expected angular-plane intermittency index for a critical system before using the CMC value as ground truth.","section":"§II and Sec. VI.3"}],"minor_comments":[{"comment":"There are several typographical errors, including ‘a extremely hot’ and ‘behaviuor’; these should be corrected before publication.","section":"Abstract and Introduction"},{"comment":"The event-averaging convention in Eq. (1) is somewhat ambiguous as written; please clarify that the factorial moments are computed per event and then averaged, or provide the standard Bialas–Peschanski expression explicitly.","section":"§II, Eq. (1)"},{"comment":"The caption contains a grammatical issue: ‘which exhibits a secondary peaks’ should read ‘which exhibit secondary peaks.’","section":"Fig. 11 caption"},{"comment":"Reference [12] contains an unresolved ‘?’ placeholder, and §VI.1 refers to ‘Section VI 2’ without a space; both should be fixed.","section":"References"},{"comment":"No statistical uncertainties are quoted for the recovered φ2 values, and Figs. 4 and 13 show no error bars; even a bootstrap or fit-uncertainty estimate would materially strengthen the comparison between the filtered mixture and the pure CMC reference.","section":"§VI.3"}],"recommendation":"major_revision","confidential_remarks":"The paper's novelty is real, but the headline recovery claim is currently an oracle-tuned benchmark result. In my view the authors can address this within the manuscript's scope by pre-specifying ε_cut or validating it on a separate sample, reconciling the two quoted pure-CMC values, and adding a null control for the particle-level filter. If those changes are made, the paper could be suitable for publication; without them, the central quantitative claim is not established."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nThe paper does something genuinely new: it runs CMC-injected critical clusters in (eta, phi) against an EPOS background and applies a two-stage TDA-ML pipeline with a track-level density filter to recover intermittency scaling. The angular-space choice is practical for LHC analyses, the Delaunay filtration and azimuthal randomization are described carefully, and the authors honestly note that the textbook phi2 = 2/3 is for (p_x, p_y), not (eta, phi). That alone makes it worth a look for anyone working on intermittency.\n\nThe soft spot is the one the stress-test flags. The central claim—recovering phi2 ~ 0.73 from the 5% mixture against a pure-CMC reference of 0.75—rests on epsilon_cut = 0.02 rad, and the paper selects that cut after inspecting which threshold reproduces the reference (Sec. VI.3, Figs. 12–13). SHAP features motivate a range, not a specific operating point, so the quantitative recovery is a fit on the benchmark, not an out-of-sample measurement. The pure-CMC index is also quoted as 0.73 in Fig. 4 and 0.75 elsewhere, with no uncertainties, and no code or data are provided. The density filter and the filtration share the same nearest-neighbor metric, so the two stages are not independent—the authors acknowledge this in their limitations.\n\nNone of this makes the pipeline useless. It makes the headline overclaimed. If the authors reframed the paper as a benchmark with an oracle-chosen threshold, and then added a pre-registered cut or a proper selection that sets epsilon_cut without looking at the reference, the central claim would be solid. As is, it's a coherent benchmark study with a known bias, not yet a robust data-driven tool for LHC data.\n\nI'd send it to peer review—it's worth having on the record with the circularity made explicit. I'd cite it if I were actively working on TDA-in-intermittency, with the caveat noted. I'd probably bring it to reading group as a case study in benchmark fitting.\n\nBest","headline":"A useful new benchmark for TDA-ML intermittency recovery in (eta, phi), but the headline recovery is tuned to the reference and needs reframing.","tokens_in":21225,"tokens_out":3899,"would_cite":false,"duration_ms":29525,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A two-stage topological machine-learning pipeline can recover the critical intermittency signal that standard factorial moments wash out in heavy-ion collisions.","keywords":["intermittency","critical fluctuations","topological data analysis","persistent homology","machine learning","heavy-ion collisions","factorial moments","QCD critical point"],"falsifier":"Run the full two-stage pipeline on pure EPOS events with no injected CMC signal, applying the same $\\varepsilon_{\\rm cut}=0.02$ rad particle filter, and measure $\\varphi_2$ from the surviving tracks; if $\\varphi_2$ rises substantially above zero, the density cut is manufacturing the power law it is supposed to recover.","tokens_in":20160,"feed_emoji":"⚛️","tokens_out":7415,"duration_ms":65833,"temperature":0.7,"pith_summary":"The paper argues that scale-invariant critical fluctuations, even when they contribute only a few percent of the final-state tracks, can be separated from an overwhelming thermal background by a two-stage pipeline that first classifies whole events and then prunes individual tracks. In the first stage, each event's $(\\eta,\\varphi)$ particle distribution is treated as a point cloud, and Betti curves derived from a Delaunay sub-level-set filtration are corrected for multiplicity bias by azimuthal randomization; a convolutional network and a boosted-tree ensemble then separate background events from events containing an injected 5% critical signal. The second stage keeps only tracks with nearest-neighbour distance below 0.02 rad, stripping away the diffuse background and leaving the dense critical clusters. On this filtered sample the normalized factorial moments resume their power-law rise, and the recovered intermittency index $\\varphi_2\\approx0.73$ matches the pure critical reference value $\\varphi_2\\approx0.75$. If this works on real collision data, it would provide a tool for probing the QCD critical point that conventional factorial-moment analysis cannot.","feed_headline":"Topological ML restores critical scaling washed out by background","feed_subtitle":"After event and track filtering, the intermittency index returns to ≈0.73 from a 5% diluted signal.","key_machinery":"The load-bearing object is the $\\Delta$-Betti curve: for each event, a Delaunay triangulation of the $(\\eta,\\varphi)$ particle cloud is filtered by nearest-neighbour distance, the counts of connected components ($\\beta_0$) and loops ($\\beta_1$) are tracked against filtration scale $\\varepsilon$, and the same curves are recomputed after randomly shuffling azimuthal angles. Subtracting the randomized baseline leaves a fingerprint of genuine spatial clustering, and the concatenated 600-dimensional $\\Delta$-Betti vector is the sole input to the classifiers. The second-stage selector is the particle-level nearest-neighbour cut $\\varepsilon_{\\rm cut}=0.02$ rad, which physically isolates the dense Lévy clusters from the thermal bulk.","core_discovery":"The central discovery is that the dilution that hides critical intermittency can be undone geometrically. Standard normalized factorial moments $F_2(M)$ of a 5% CMC-injected mixture collapse to the inert background value $\\varphi_2\\approx0$, because the non-critical tracks outnumber the critical clusters. After an event-level topological classifier retains only high-score events, and after a particle-level density cut $d_{\\rm NN}\\le0.02$ rad removes the diffuse thermal bulk, the same factorial moments recover a power law and give $\\varphi_2\\approx0.73$, consistent with the pure CMC reference $\\varphi_2\\approx0.75$. The paper also establishes a calibration point for this geometry: the pure CMC value in $(\\eta,\\varphi)$ space is about 0.75, shifted from the theoretical $2/3$, which the paper notes is a prediction for $(p_x,p_y)$ space, not angular space.","pith_inferences":["Editorial inference: the choice of $\\varepsilon_{\\rm cut}=0.02$ rad was validated against the known pure CMC reference, so applying this pipeline to real data, where no pure reference exists, will require an independent calibration of the density threshold or the method risks circularity.","Editorial inference: because the paper itself notes that the theoretical $\\varphi_2=2/3$ applies to $(p_x,p_y)$ and not $(\\eta,\\varphi)$, the recovered values 0.73 and 0.75 are geometric calibration numbers; testing the Ising prediction directly would require running the same $\\Delta$-Betti pipeline in a cumulatively flattened momentum space.","Editorial inference: a sharp testable extension is to run the second-stage density filter on pure EPOS events and on a Poissonian toy background with no injected clusters; if $\\varphi_2$ rises substantially above zero, the filter itself would be creating apparent intermittency.","Editorial inference: the nearest-neighbour threshold sits exactly in the angular-separation regime where detector track merging and splitting occur, so simulating a realistic detector response is the immediate next step before this can be applied to experimental data."],"forward_implications":["At a 5% injected signal fraction, event-level topological classification reaches an AUC of about 0.99 with both a convolutional network and a boosted-tree ensemble, while performance degrades to about 0.95 at 3% and about 0.8 at 1%.","Factorial moments computed on event-filtered samples stay flat, so the particle-level density cut is both necessary and sufficient to restore the critical power-law scaling.","Applying the density cut to the pure CMC sample does not distort its scaling, indicating that the recovered index reflects the injected signal rather than an artifact of the filter.","The pure CMC baseline in $(\\eta,\\varphi)$ space is $\\varphi_2\\approx0.75$, so future intermittency searches in angular coordinates should compare against this shifted reference rather than the $(p_x,p_y)$ value of $2/3$.","The same pipeline can be recalibrated with dynamically generated critical configurations, such as the successive-contraction-and-randomization model, to test whether the recovery survives more realistic phase evolution."],"supporting_citations":[{"why":"Defines normalized factorial moments and the intermittency power law that the pipeline aims to restore.","marker":"[15, 17]"},{"why":"Supplies the Critical Monte Carlo Lévy-walk algorithm that generates the scale-invariant critical clusters used as the injected signal.","marker":"[31]"},{"why":"Formulates intermittency analysis in the (η, φ) plane, the framework this work extends.","marker":"[32]"},{"why":"Provides the earlier toy-model study of factorial moments in (η, φ) space that the 5% mixture's scaling behaviour resembles.","marker":"[38]"},{"why":"Supplies the EPOS background events into which the CMC signal is embedded and the baseline the classifiers must reject.","marker":"[39–42]"},{"why":"Provides the Delaunay triangulation, phantom-point azimuthal boundary handling, and Betti-curve construction for nuclear-collision point clouds.","marker":"[49]"},{"why":"Introduces the azimuthal randomization baseline used to define the dynamical Δ-Betti curves.","marker":"[50]"},{"why":"Supplies the topological machine-learning approach and the convolutional architecture for classifying weak intermittency signals.","marker":"[51]"},{"why":"Provides the gradient-boosted decision-tree implementation used as the second, interpretable classifier.","marker":"[53]"},{"why":"Provides the feature-attribution method used to identify which filtration scales drive classification.","marker":"[55]"}],"fun_headline_variants":["Topological ML recovers intermittency index 0.73 from 5% signal","Two-stage ML filter exposes QCD critical fluctuations","Topology+ML recovers critical scaling from 5% background","Geometry plus deep learning recovers hidden intermittency in heavy-ion data","Topological data analysis + AI probes QCD critical point"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The recovery rests on the assumption that the 0.02-radian density threshold can be fixed from the pure CMC reference without peeking at the injected signal, and that a Lévy walk with exponent 1/6 performed directly in (η, φ) is the right model for critical angular fluctuations.","fun_headline_variants_meta":{"raw":{"variants":["Topological ML recovers intermittency index 0.73 from 5% signal","Two-stage ML filter exposes QCD critical fluctuations","Topology+ML recovers critical scaling from 5% background","Geometry plus deep learning recovers hidden intermittency in heavy-ion data","Topological data analysis + AI probes QCD critical point"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001725,"raw_usage":{"total_tokens":6864,"prompt_tokens":1028,"completion_tokens":5836,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":644,"completion_tokens_details":{"reasoning_tokens":5747}},"tokens_in":644,"tokens_out":5836,"duration_ms":39112,"temperature":1.0,"reasoning_tokens":5747,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:12:43.780381+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the full two-stage pipeline on pure EPOS events with no injected CMC signal, applying the same $\\varepsilon_{\\rm cut}=0.02$ rad particle filter, and measure $\\varphi_2$ from the surviving tracks; if $\\varphi_2$ rises substantially above zero, the density cut is manufacturing the power law it is supposed to recover.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the Critical Monte Carlo Lévy-walk algorithm that generates the scale-invariant critical clusters used as the injected signal."},{"cited_title":"NA61/SHINE overview","cited_arxiv_id":"2402.10973","evidence_quote":"Formulates intermittency analysis in the (η, φ) plane, the framework this work extends."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the earlier toy-model study of factorial moments in (η, φ) space that the 5% mixture's scaling behaviour resembles."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the feature-attribution method used to identify which filtration scales drive classification."}],"review_version":1}